TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Cloudflare says it has released two decision models, Clef and Clef-flash, on Workers AI and under an Apache 2.0 license on Hugging Face. The company also introduced a reinforcement-learning product for customers to fine-tune Clef; independent verification of the company’s benchmark claims and further product details were not included in the announcement.
Cloudflare has released Clef and Clef-flash, two decision models hosted on its Workers AI platform and published on Hugging Face under an Apache 2.0 license. The company also announced a new reinforcement-learning (RL) product that lets customers fine-tune Clef for their own use cases, putting model customization alongside the open-source release.
Cloudflare describes decision models as systems that return structured classifications and probabilities for use in software workflows. For example, an application could send a customer support message and receive an urgency estimate and a suggested team, then use those outputs to route the ticket, escalate it or send it to a person. The company presents this approach as distinct from general-purpose large language models, which can generate open-ended text and tool calls but are less deterministic.
The company says both Clef models are compatible with the Jev API and can be tried through Workers AI or run locally from their Hugging Face release. Cloudflare describes Clef as having a vision encoder for classifying images and a 64,000-token context window; it contrasts those features with Jev’s text-only classification and 32,000-token context window, as described in the post.
Cloudflare’s post reports benchmark results across decision-making and classification tasks, including tool retrieval, API use and phishing classification. It says Clef led the Jev Decision Index evaluation cited in the announcement, while Clef-flash had particularly strong results on some tasks. These are company-reported evaluations; the post points readers to a live benchmark demo but the supplied material does not include independent assessments.
Structured Decisions for AI Workflows
Decision models are intended to turn model outputs into typed results that software can act on, rather than requiring an application to interpret free-form text. That can be useful in workflows such as support routing, domain categorization and incident triage, where a system needs to choose among defined outcomes and may need to pass uncertain cases to a human.
Cloudflare’s release combines three pieces: models customers can access through its hosted AI service, weights available under a permissive open-source license, and an announced route to customize Clef through RL fine-tuning. If the tools perform as described, developers could test a decision model locally, deploy it on Cloudflare’s infrastructure or adapt it to a particular task. The announcement does not establish that these models are suitable for autonomous decisions in every setting; users would still need to evaluate errors, confidence thresholds and when human review is necessary.
As an affiliate, we earn on qualifying purchases.
Cloudflare’s Decision Model Comparison
Cloudflare frames Clef as an alternative in a growing category of decision-focused models, and compares it with Typesafe AI’s Jev System One and other models in its published evaluations. Its stated goal is different from replacing general-purpose language models: Clef is designed for bounded decisions within a workflow, while an LLM can generate broader responses or use tools.
In one internal example, Cloudflare says its Threat Intelligence team used Clef with Browser Run to fetch and render a website, then classify its domain. The company reports that this workflow took 2.2 seconds and returned several classifications with probabilities. It compares that result with 4.7 seconds for gpt-oss-120b in the same workflow, which Cloudflare says returned two classifications. These figures describe the company’s test and are not a general performance guarantee.
Cloudflare also attributes low network latency to hosting the models on Workers AI and its edge infrastructure. Its post includes median and 95th-percentile latency figures from its evaluations, but those results depend on the tested models, tasks and setup.
“Clef is currently the leader when evaluated against the Jev Decision Index.”
— Cloudflare, in its announcement
Open Questions on Fine-Tuning
The supplied announcement does not specify the RL product’s availability, pricing, technical workflow or access requirements. It also does not explain what training data customers must provide, what controls are available during fine-tuning, or how a customized model can be evaluated and deployed.
Cloudflare’s performance and quality comparisons come from the company’s own tests. The source material does not describe independent replication, detailed benchmark methodology or results across all real-world deployments. It is also unclear how Clef handles ambiguous inputs and what confidence threshold Cloudflare recommends before an application acts without human review.
Availability and Evaluation Ahead
Developers can look to Workers AI and Hugging Face to access the hosted models and open-source releases described by Cloudflare. The company’s live decision-index demo is cited as a place to review benchmark results. Further product information will be needed to establish when and how customers can use the RL fine-tuning service.
For teams considering adoption, the next practical step is to test Clef on their own inputs and compare its decisions, latency and failure cases with existing systems. Cloudflare has not stated a date for additional releases or supplied a schedule for expanded fine-tuning features in the source material.
Key Questions
What did Cloudflare announce?
Cloudflare announced Clef and Clef-flash, decision models hosted on Workers AI and released on Hugging Face under an Apache 2.0 license, along with an RL product for fine-tuning Clef.
What does a decision model do?
It returns structured classifications and probabilities that an application can use to route work, trigger an action or send a case to a human for review.
Can developers run Clef outside Workers AI?
Cloudflare says the models are available on Hugging Face under Apache 2.0, allowing developers to run them locally and experiment with them.
How does Clef compare with Jev?
Cloudflare reports that Clef performed competitively or led on several of its listed evaluations, and says it has image-classification capability and a 64,000-token context window. Those comparisons are company-reported, and outcomes vary by benchmark.
What is known about the RL fine-tuning product?
Cloudflare says the product will let customers fine-tune Clef for their use cases. The announcement does not give details on availability, pricing, training requirements or the fine-tuning process.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
